International Journal of Advanced Mechatronic Systems (9 papers in press)

Regular Issues

• Classification of malware family in large executable files using NeASA-Net in MapReduce framework  
  by Manoj D. Shelar, S. Srinivasa Rao 
  Abstract: This paper proposes the neuron attention stacked autoencoder (NeASA-Net) to classify the malware family from the executable files. The classification process is done in the MapReduce framework. At first, the accumulated input executable files are subjected to the mapper phase. Here, the features, like opcode 4-gram, API 4-gram, file size, and PE section size are determined. Then, the determined features are merged and subjected to malware family classification using NeASA-Net in the reducer phase. The NeASA-Net is introduced by combining deep stacked autoencoder (DSA) with a neuron attention stage-by-stage net (NASNet). Malware is finally classified as Gatak, Tracur, Obfuscator.ACY, Simda, Kelihos_ver1, Vundo, Lollipop, ramnit, and Kelihos_ver3. The performance of the NeASA-Net model is validated by comparing it with traditional detection models. Here, the NeASA-Net model achieved superior performance with an accuracy of 92.77%, a true positive rate (TPR) of 95.98%, and a false positive rate (FPR) of 6.54%.
  Keywords: neuron attention stacked autoencoder; NeASA-Net; deep stacked autoencoder; DSA; fractional calculus; neuron attention stage-by-stage net; NASNet; cyber security.
  DOI: 10.1504/IJAMECHS.2025.10071223
   
  
• Enhancing the efficiency of 4D printing design through time-series prediction by dense-LSTM crossed network  
  by Yifan Xu, Mengtao Wang, Hidemitsu Furukawa, Zhongkui Wang, Qi Li, Lin Meng 
  Abstract: This paper proposes a deep learning-based approach to predict the deformation of 4D-printed hydrogel models with varying lengths, aiming to improve the efficiency of the design process. A voxel-based modeling method is used to create hydrogel models in Abaqus, and their deformation data is obtained through finite element analysis (FEA). A mixed dataset is then constructed by mapping each models expansion rate sequence to its corresponding deformation outcome. Based on this dataset, a novel deep learning architecture called dense-LSTM crossed network (DSCN) is introduced and trained. The trained model enables direct prediction of deformation results from the initial model parameters, reducing reliance on time-consuming simulation. Experimental results show that the proposed method shortens the model design verification process by 1.52%, thus enhancing the overall efficiency of 4D printing design. This study demonstrates the potential of combining intelligent modeling with deep learning to streamline additive manufacturing workflows involving shape-morphing materials.
  Keywords: 4D printing; deep learning; hydrogel; recurrent neural network; RNN; voxelisation design; mixed dataset.
  DOI: 10.1504/IJAMECHS.2025.10071224
   
  
• A comprehensive review of inverse kinematics techniques from analytical foundations to artificial intelligence integration  
  by Navya Manjegowda, Muralidhara, Nirmith R. Jain, Sandeep Kumar Shivaswamy 
  Abstract: Inverse kinematics (IK) is crucial for enabling precise and coordinated motion in robotic systems. This review provides a comprehensive analysis of IK techniques in robotics, examining the traditional analytical, numerical methods and the transformative impact of artificial intelligence (AI)-based approaches. It explains IK fundamentals such as forward kinematics (FK) and joint angles, emphasising the analytical clarity of traditional methods and the adaptability of numerical IK. The evolution of AI-based IK, using machine learning (ML) and neural networks (NNs), is highlighted for its versatility and optimisation in various robotic applications. A comparative analysis outlines the strengths and limitations of traditional and AI-based IK methods, highlighting their efficiency in different scenarios. As robotics advances, the fusion of these IK techniques becomes crucial for navigating complexities and enhancing capabilities. The review underscores the symbiotic integration of classical, numerical, and AI-based IK techniques to meet the demands of modern robotics, fostering improved performance.
  Keywords: robotics; inverse kinematics; traditional methods; machine learning; neural networks; optimisation; motion planning.
  DOI: 10.1504/IJAMECHS.2025.10071298
   
  
• Optimising production efficiency in small and medium enterprises with IoT-driven automated jerry can filling systems  
  by Anshu Prakash Murdan, Dhanveer Ranjus 
  Abstract: This paper presents a cost-effective, IoT-based automated filling and capping system tailored specifically for small and medium-sized enterprises (SMEs). The solution leverages open-source Arduino controllers, modular hardware, and advanced sensor integration, including ultrasonic and infrared sensors, to enable precise, real-time operational adjustments. While traditional automation technologies such as PLCs and SCADA remain widely used in industries, this system offers a more affordable alternative, providing superior scalability, flexibility, and ease of maintenance. Comprehensive operational and field testing, supported by statistical analyses, demonstrated substantial improvements including a 60% increase in throughput, an 85% improvement in filling accuracy, and halving material waste and downtime. Economic analysis further validated the financial advantages, highlighting a rapid return on investment. Operator feedback emphasised its reliability and intuitive interface. By integrating real-time monitoring, adaptive control algorithms, and predictive maintenance capabilities, the developed system establishes a new benchmark for accessible, efficient, and scalable automation solutions suitable for SMEs.
  Keywords: automation; internet of things; IoT; prototype; real-time monitoring; sensor calibration; cost-effectiveness.
  DOI: 10.1504/IJAMECHS.2025.10071770
   
  
• Autism spectrum disorder detection using convolutional neural network with transfer learning  
  by Rakhee Kundu, Sunil Kumar 
  Abstract: Autism spectrum disorder (ASD) is a complex neurodevelopmental condition, which can vary widely among individuals, making it difficult to establish a uniform set of criteria for diagnosis, leads to underdiagnosis or misdiagnosis in previous researches. This research developed the green anaconda one-to-one-based optimiser-based convolutional neural network with transfer learning (GAOOBO_CNN_TL) for ASD classification utilising multimodal data. Firstly, the image pre-processing process is performed by Kuwahara filters and RoI extraction. Then, the input autism data is normalised by using Z-score. Then, the selection of features is conducted by the Hubert index and data augmentation is done by employing bootstrapping. Then, the ASD classification is done by CNN_TL. Then, weight optimisation is executed using GAOOBO, which is the incorporation of green anaconda optimisation (GAO) and one-to-one-based optimiser (OOBO). The GAOOBO_CNN_TL gained accuracy with 92.823%, specificity with 93.127% and sensitivity with 91.997%.
  Keywords: autism spectrum disorder; ASD; green anaconda optimisation; GAO; GoogLeNet; Xception; convolutional neural network; CNN.
  DOI: 10.1504/IJAMECHS.2025.10071087
   
  
• Credit card fraud detection using hybrid Chord-SpinalNet  
  by Muzaffar Abdur Rahim Shabad, M. Kavitha 
  Abstract: Due to the advancement of e-commerce systems and information technologies, credit cards act as the popular payment method for both normal and online shopping. Still, fraudulent transactions are a critical issue for online payments and credit card fraud has become more prevalent in recent years. This paper develops a hybrid network-based credit card fraud detection (CCFD) technique, named Chord-SpinalNet by merging the Chord layer with SpinalNet. The credit card fraud detection dataset is used for the evaluation. This data covers more features, and the feature fusion is done by the Jensen difference for feature sorting and the deep neuro fuzzy network (DNFN) for generating fusing coefficients. The Chord-SpinalNet is utilised in the CCFD phase to identify unauthorised transactions. The accuracy, sensitivity and specificity metrics are utilised to determine the performance of the Chord-SpinalNet and it obtained the highest accuracy, sensitivity and specificity of 0.923, 0.915 and 0.926.
  Keywords: credit card fraud detection; SpinalNet; Jensen difference; Chord layer; deep neuro fuzzy network.
  DOI: 10.1504/IJAMECHS.2025.10071090
   
  
• Optimal light gradient boosting model-based prognostic approach for remaining useful battery life prediction  
  by Bighnaraj Naik, Geetanjali Bhoi, Rajat Kumar Sahu, V. Ashok Gajapathi Raju, Janmenjoy Nayak, Manohar Mishra 
  Abstract: Predicting battery life in high-power applications is crucial for ensuring uninterrupted operations in fields like electric vehicles, aerospace, and renewable energy storage. Accurate life predictions enable proactive maintenance, reducing unexpected failures and downtime. In safety-critical sectors such as medicine and defence, timely battery replacement prevents malfunctions and ensures reliability. Effective battery management extends lifespan, minimises waste, and supports environmental sustainability through proper disposal and recycling. In smart energy systems, forecasting battery life enhances grid stability, load balancing, and renewable energy integration. This study employs a light gradient boosting model (LGBM) to predict the remaining useful battery life based on voltage and current behaviour. The model's performance is improved through optimised hyperparameters using grid and randomised searches. Addressing challenges like complex battery behaviour and variable conditions, the proposed approach is compared with state-of-the-art models, demonstrating competitive performance in remaining useful life (RUL) prediction.
  Keywords: battery remaining useful life; sustainability; safety; machine learning; light gradient boosting machine; LGBM; hyperparameter optimisation.
  DOI: 10.1504/IJAMECHS.2025.10071089
   
  
• Empirical approach for a novel PCC-MFCC and TS-CTRNN based speech recognition system  
  by Shivani Trivedi, Sanjay Patidar, Rohit Rastogi 
  Abstract: Currently, in the field of speech signal processing, a large amount of research has been conducted. Especially, there is a growing interest in the automatic speech recognition (ASR) technology field. Nevertheless, owing to a noisy environment, traditional systems have low performance. Hence, a novel Tanh sigmoid-centred continuous time recurrent neural network (TS-CTRNN)-cantered speech recognition system (SRS) is proposed in this work. Two phases are incorporated by the proposed technique. Firstly, in the input audio, the frequency spectrum is scrutinised. Next, the spectrum is pre-processed. Afterwards, from the pre-processed signal, the features are extracted. The next phase begins with pre-processing and word embedding where the label is taken as the input. At last, the output obtained from both phases is inputted into the TS-CTRNN, which predicts speech in the format of text. The experimental outcomes exhibit that when analogised to the created ASR system, the enhanced virtue of noise elimination methodology and TS-CTRNN-cantered recognition provides a better relative enhancement of accuracy to (96.89%).
  Keywords: Pearson correlation coefficient based Mel frequency cepstral coefficient; PCC-MFCC; entropy-based Wiener filter; Tanh sigmoid based continuous time recurrent neural network; TS-CTRNN; XLM-RoBERTa; automatic speech recognition; ASR.
  DOI: 10.1504/IJAMECHS.2025.10071088
   
  
• Object detection for advanced driving assistance system using a novel machine learning approach: rat swarm Henry gas solubility trained deep quantum network  
  by Babruvan Ramrao Solunke, Sachin Ratikant Gengaje 
  Abstract: Due to the unique technologies of intellectual transportation model, the intelligent vehicle becomes the carrier for inclusive amalgamation of several technologies. Even though, vision-assisted automated driving has acquired effective prospects, there becomes an issue to evaluate the complex traffic situations due to accumulated data. Thus, a deep model driven hybridised optimisation is needed for detecting multi-objects. The input video frame first undergoes pre-processing using a Laplace filter to eliminate noise. The pre-processed images are fed to object segmentation using fuzzy local information C-means (FLICM). The segmented images are allowed for feature extraction. For multi-object detection, the extracted features are processed by the deep quantum neural network (DQNN), which is optimised through the rat swarm Henry gas solubility optimisation (RSHGSO) technique. The RSHGSO is the incorporation of rat swarm optimisation (RSO) and Henry gas solubility optimisation (HGSO). RSHGSO-DQNN has significantly lower processing time compared to baseline model. The HGSO-DQNN demonstrated enhanced performance, achieving a remarkable accuracy of 90.9%, an F-measure of 94.7%, and a precision of 92.6%.
  Keywords: multi-object detection; deep quantum neural network; DQNN; Laplacian filter; FLICM; image processing.
  DOI: 10.1504/IJAMECHS.2025.10071091